A new framework for analyzing color models with generative adversarial networks for improved steganography

Steganography algorithms are designed by human experts and use rule-based methods to hide the secret data inside a cover medium. Recently, this process has been automated by using Generative Adversarial Networks. However, Generative Adversarial Networks based steganography has certain limitations. These algorithms may not generate steganographic images that are real enough to fool the intruder. Security of steganography degrades as embedding capacity is increased. A new framework for improved steganography based on deep learning is proposed in this paper. The new framework uses color models with Generative Adversarial Networks. Popular color models such as HSV, YCrCb, YDbDr, CIE-XYZ, YIQ, HED, and YUV are analyzed. The proposed framework helps to identify the best color model with Generative Adversarial Networks based steganography. The quality of steganographic images is assessed with three metrics: capacity, security, and invisibility. Extensive experimentation has been carried out with the CelebA dataset which has more than 200 K samples to assess the effectiveness of the proposed framework. The results show that compared with the RGB and other color models, the CIE-XYZ model produces the best results with Generative Adversarial Networks for steganography.The results further show that the proposed framework improves the security of steganography even when the embedding capacity is increased.